{"id":"4cb2f3f4-e42c-42b0-a0d1-c9a00f71f41d","arxiv_id":"2501.18862","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Local and cluster distributed reproduction numbers, built from the next-generation matrix and infection ratios, plus a bounded-Gaussian and shuffle privacy mechanism, give accurate private estimates of regional epidemic spread in COVID-19 data.","lead":"The paper turns one network-wide epidemic reproduction number into per-link and per-cluster numbers, then wraps them in a differential-privacy mechanism so they can be shared without revealing travel patterns. This gives public-health analysts a finer-grained, formally private view of where an outbreak is growing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Privacy guarantee rests on an unvalidated sensitivity bound: k=1e-5 is asserted, not derived, and the DP claim fails if a single mobile record can move ζ_i by more than k.","rationale":"The reader identified the same load-bearing concern: the asserted adjacency parameter k=1e-5 in Section V-C is the linchpin of the differential-privacy analysis, and it is not derived. My stress-test confirms that this is the weakest point in the central claim. The threshold theorems (Theorems 1 and 3) are straightforward algebra under x(t) >> 0 and are internally consistent; the cluster aggregation formulas in Definition 11 and Corollary 2 are coherent. The DP mechanism itself (bounded Gaussian + shuffler) is a standard import from prior work, and the accuracy analysis in Theorem 4 is a routine moment calculation. However, the formal privacy guarantee applies only to adjacent vectors within l2 distance k, and the paper never shows that replacing one individual's mobility record changes a local authority's ζ_i by at most k. The single sentence in Section V-C is the entire support for this. Because the reader's verdict was already CONDITIONAL with the explicit condition of supplying a rigorous sensitivity bound, my finding does not shift the verdict. A secondary concern, also noted by the reader, is that the abstract overclaims 'new conditions for the existence, uniqueness, and stability of equilibrium states' that never appear in the theorems; this is a presentation issue rather than the core technical vulnerability. With the sensitivity bound supplied and the overclaim corrected, the framework could stand as a solid extension of the authors' earlier ACC work.","tokens_in":31119,"tokens_out":4816,"duration_ms":51956,"concrete_test":"Compute the actual l2 sensitivity of ζ_i with respect to a single mobile data point. Using the same SafeGraph data and the transmission-rate construction from [16], for each local authority i, select each individual record (or, at the CBG aggregation level, each device) and perturb its mobility pattern by its maximum possible change (e.g., reassigning its home CBG or altering its dwell time). Recompute β_ij, then ζ_i, and record max ||ζ_i - ζ_i'||_2 over all i and all single-record perturbations. If this maximum exceeds k=1e-5, the claimed differential-privacy guarantee is invalid. Alternatively, derive an analytic bound from the formula for β_ij: if β_ij = c·flow_ij/N_i, then a single device changes flow_ij by at most 1, bounding the change in each entry of ζ_i; compare the resulting l2 bound to k.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The entire differential-privacy contribution is calibrated to the l2 sensitivity of the identity mapping on local aggregated effective reproduction number vectors ζ_i. In Definition 13 and Remark 11, the paper defines the sensitivity as the maximum over adjacent vectors ζ_i ∼ ζ'_i, and then observes that the identity mapping has sensitivity exactly k. But this is tautological: the adjacency relation in Definition 3 is defined by the l2-distance bound k, so the sensitivity of the identity mapping over that adjacency is k by construction. The actual load-bearing question is whether any two local aggregated ERN vectors that arise from real neighboring mobility databases (i.e., databases differing in a single individual's record) are within l2 distance k. Section V-C asserts that k=1e-5 is 'chosen by the maximum variation in the distributed ERNs that a single mobile data point can cause when it changes by its maximum possible amount,' but no derivation, formula for β_ij from SafeGraph, or code is provided. The mechanism's noise variance in (15) scales with k, and the shuffle amplification in Lemma 2 builds on the per-vector ε0 guarantee. If the true per-record sensitivity exceeds k, the noise is under-calibrated and the claimed (ε0, δ)-differential privacy (and the subsequent amplification) does not hold for the real data. This is not an internal inconsistency; it is an unverified external calibration that is load-bearing for the paper's core privacy claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines local distributed effective reproduction numbers (LERNs) for each node of a network SIS/SIR model, proves that each LERN exceeds one exactly when the node's infected proportion is increasing, and extends this to cluster-level effective reproduction numbers (CERNs) that threshold the summed infection rate in a cluster. It then develops a differential privacy framework in which local authorities add bounded Gaussian noise to local aggregated effective reproduction number vectors, a shuffler amplifies privacy, and central aggregators form privatized cluster-level reproduction numbers. The framework is validated on SafeGraph mobility data for 1,023 US regions grouped into 100 clusters.","tokens_in":31302,"tokens_out":9609,"duration_ms":91250,"significance":"If its claims hold, the paper contributes a scalable, interpretable monitoring tool: each region or cluster can share a scalar threshold signal without releasing raw mobility flows, and the local-to-cluster aggregation is exact. The privacy framework is a sensible application of existing mechanisms (bounded Gaussian noise plus shuffling), and the empirical evaluation on real mobility data is a strength. However, the central threshold property is largely a rearrangement of the model equations: since \\bar R_i^t = (\\dot x_i + \\gamma_i x_i)/(\\gamma_i x_i), the statement that \\bar R_i^t > 1 iff \\dot x_i > 0 is immediate from the definition. The paper also overclaims in the abstract by promising conditions for existence, uniqueness, and stability of equilibria, which do not appear in the body. Most importantly, the differential privacy guarantee depends on an asserted, not derived, sensitivity bound for real mobility data, and the main accuracy theorem contains an apparent algebraic error.","major_comments":[{"comment":"The differential privacy guarantee is calibrated to the adjacency parameter k = 10^{-5}, asserted in Section V-C as \"the maximum variation in the distributed ERNs that a single mobile data point can cause when it changes by its maximum possible amount.\" No derivation, formula, or code is provided for this bound. Because Definition 3 defines adjacency by \\|\\zeta - \\zeta'\\|_2 \\le k, the sensitivity of the identity mapping is k by construction (Remark 11), so the load-bearing question is whether any two local aggregated ERN vectors arising from real neighboring mobility databases are within l2 distance k. If a single mobile record can change \\zeta_i by more than k, the noise variance in (15) is under-calibrated and the claimed (\\epsilon_0,\\delta)-differential privacy, and hence the shuffle amplification in Lemma 2, does not hold for the actual data. Please provide a derivation from the transmission-rate formula in [16] or an explicit conservative upper bound.","section":"Section V-C and Definition 13 / Remark 11"},{"comment":"The stated first and second moments of the private cluster distributed effective reproduction numbers are inconsistent with their definition in (14). According to (14), \\tilde R^t_{\\chi_q,\\chi_r} = (\\sum_{k\\in\\chi_q} \\tilde R^t_{k,\\chi_r})/(\\sum_{k\\in\\chi_q} \\gamma_k x_k(t)), so its expectation should be a single weighted average over i \\in \\chi_q with denominator \\sum_{i\\in\\chi_q} \\gamma_i x_i(t). The formula given in Theorem 4 instead sums over all clusters q=1,\\dots,m in the numerator and divides by \\sum_{i\\in(\\cup\\chi_q)} \\gamma_i x_i(t), which appears to be an algebraic error. This is load-bearing because Theorem 4 is the formal accuracy guarantee for Problem 6.","section":"Section IV-D, Theorem 4"},{"comment":"The abstract states that the derived conditions are \"used to derive new conditions for the existence, uniqueness, and stability of equilibrium states of the underlying epidemic model,\" but the manuscript contains no such theorem. Theorem 2 and Corollary 1 relate the local effective reproduction numbers to the spectral radius of the next-generation matrix, which is a threshold condition for the global reproduction number, not an existence, uniqueness, or stability result for equilibria. The introduction's contribution list repeats this claim. The abstract and contributions should be revised to match the actual content, or the missing analysis should be added.","section":"Abstract and Section I"}],"minor_comments":[{"comment":"The last row of the cluster distributed effective reproduction number matrix shows \\bar R^t_{\\chi_m,\\chi_1} twice; the final entry should be \\bar R^t_{\\chi_m,\\chi_m}.","section":"Definition 12, equation (9)"},{"comment":"The numerator is written as \\sum_{k\\in\\chi_q} \\tilde R^k_{,\\chi_r}, which appears to be a typo for \\sum_{k\\in\\chi_q} \\tilde R^t_{k,\\chi_r}.","section":"Equation (14)"},{"comment":"The sentence \"Consider |\\chi_q| = m \\le n clusters\" uses m for both the size of the cluster and the total number of clusters; please use |\\chi_q| for the cardinality to avoid confusion.","section":"Proof of Theorem 3"},{"comment":"The notation \\tilde R_{\\chi_q,\\chi_r} in the theorem drops the time superscript t used elsewhere; please keep notation consistent.","section":"Theorem 4 statement"},{"comment":"The phrase \"capturing causal relationships\" is too strong for the described empirical analysis, which is correlational; please rephrase.","section":"Section V-B"},{"comment":"The RMSE values are based on 100 samples with no confidence intervals; as the percentage errors are small, reporting standard errors or confidence intervals would strengthen the accuracy claim.","section":"Section V-C, Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable control-systems application paper, and the aggregation hierarchy and privacy architecture are useful. However, the advertised existence/uniqueness/stability contribution is absent, the sensitivity calibration for the privacy guarantee is unsupported, and Theorem 4 appears to contain an algebraic error. These issues are fixable within the scope of the paper, but they are load-bearing for the main claims and require a substantive revision rather than a minor edit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this paper. The genuinely new pieces are the cluster-level aggregation rules (Definition 12, Corollary 2) and the differential privacy pipeline that lets local authorities report privatized cluster effective reproduction numbers. The hierarchical aggregation formula is clean, and the real-data experiments do show cluster metrics tracking regional infection waves better than a single network-level Rt. That part is useful and coherent.\n\nWhat is not new: the local distributed BRN matrix is exactly the next-generation matrix Gamma^{-1}B, which the paper acknowledges in Remark 3. And Theorem 1 (local ERN > 1 iff x_i is increasing) is true by construction: the LERN is defined as infection inflow divided by gamma_i x_i, so the threshold statement is an algebraic restatement of the model. Theorem 3 is the same at cluster level. Theorem 2 is a reasonable observation using a similarity transform to a row-stochastic matrix, but it is a small step. So the mathematical payload is largely forced by the definitions.\n\nThe bigger problems are two. First, the abstract claims \"new conditions for existence, uniqueness, and stability of equilibrium states,\" but the body never derives any such conditions. That overclaim needs to go. Second, the DP guarantee is calibrated to the adjacency parameter k=1e-5 in Section V-C. As written, the sensitivity of the identity mapping equals k by definition of adjacency (Definitions 3 and 13), so the mechanism is internally consistent. But the load-bearing question is whether real neighboring mobility databases produce local aggregated ERN vectors within l2 distance k. The paper asserts this with no derivation, no formula for beta_ij from SafeGraph data, and no code. If a single mobile record can move zeta_i by more than k, the noise is under-calibrated and the privacy claim fails for real data. That uncertainty is the main barrier to trusting the privacy contribution. The [0,14] cap and the forced positive infections in data processing also shape the reported metrics; they are disclosed, but they deserve more scrutiny.\n\nWho should engage with this? Epidemic modelers and control engineers who want interpretable, multi-scale spread indicators with a formal privacy layer. It deserves a serious referee, but the referee should focus on the sensitivity calibration and the abstract overclaim. With those fixed, the framework is a modest but solid extension of the authors' earlier ACC work.\n\nRecommendation: send to peer review, with the sensitivity derivation and abstract as required revisions.","headline":"Useful cluster-level aggregation and a coherent DP pipeline, but the privacy guarantee rests on an asserted sensitivity k and the abstract overpromises equilibrium results the body never proves.","tokens_in":31913,"tokens_out":3359,"would_cite":false,"duration_ms":32464,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92D30","68P27"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper defines a local effective reproduction number for each node whose threshold at 1 exactly tells whether that node's infection level is increasing, decreasing, or flat, and shows the same threshold aggregates to clusters and…","keywords":["distributed reproduction numbers","network epidemic models","effective reproduction number","cluster aggregation","differential privacy","bounded Gaussian mechanism","shuffle model","mobility data"],"falsifier":"Recompute the local aggregated ERN vector $\\zeta_i$ after adding and after removing one individual's daily visits in the Section V mobility data, using the same transmission-rate estimator; if the largest $\\ell^2$ change exceeds $k=10^{-5}$, the differential-privacy guarantee claimed for interaction frequencies does not hold on that data.","tokens_in":30835,"feed_emoji":"🦠","tokens_out":11629,"duration_ms":109393,"temperature":0.7,"pith_summary":"Network-level reproduction numbers summarize the whole epidemic but hide where it is spreading. This paper proposes a per-node local effective reproduction number, $\\bar R^t_i = \\sum_j s_i(t)\\beta_{ij}x_j(t)/(\\gamma_i x_i(t))$, and proves that comparing it to 1 gives an exact, real-time sign detector for that node's infection: above 1 means rising, below 1 means falling, equal means flat. It then defines cluster-level versions that aggregate from local numbers at any scale, so regional and national monitors can use the same threshold logic. Because these numbers encode how often people interact, the paper adds a privacy layer: each local authority perturbs its aggregated vector with a bounded Gaussian mechanism, a shuffler anonymizes the vectors inside each cluster, and the resulting cluster numbers carry quantified first- and second-moment accuracy. A COVID-19 mobility example suggests the privatized cluster numbers remain useful, with roughly 5 to 9 percent average error at strong privacy.","feed_headline":"Local epidemic thresholds hold under differential privacy","feed_subtitle":"A per-node reproduction number signals rising or falling infection; bounded noise and shuffling keep raw mobility data hidden.","key_machinery":"The carrying object is the local distributed effective reproduction number vector, whose ith row has entries $\\bar R^t_{ij}=s_i(t)\\beta_{ij}x_j(t)/(\\gamma_i x_i(t))$; its row sum is the local ERN $\\bar R^t_i$. This vector is a diagonal similarity transform of the next-generation-matrix-based pseudo-ERN matrix, so its spectral radius matches the network-level reproduction number while its row sums encode local thresholds. For privacy, the machinery is the bounded Gaussian mechanism applied to the pre-aggregated local vectors $\\zeta_i$, with the shuffler model amplifying the local $\\epsilon_0$-differential privacy to central $(\\epsilon,\\delta)$-differential privacy, and Theorem 4 gives the accuracy of the resulting private cluster distributed ERNs.","core_discovery":"The paper's central claim is that the sign of $\\dot x_i(t)$ is fully determined by the local effective reproduction number $\\bar R^t_i$, the row sum of the local distributed ERN matrix whose entries are $\\bar R^t_{ij}=s_i(t)\\beta_{ij}x_j(t)/(\\gamma_i x_i(t))$. Theorem 1 states $\\bar R^t_i>1$ iff $x_i$ is increasing, $\\bar R^t_i<1$ iff it is decreasing, and $\\bar R^t_i=1$ iff it is flat, provided all infection proportions are positive. The same construction is lifted to clusters: the cluster effective reproduction number $\\bar R^t_{\\chi_q}=\\sum_{i\\in\\chi_q}\\gamma_i x_i\\bar R^t_i/\\sum_{i\\in\\chi_q}\\gamma_i x_i$ has the corresponding threshold for the sum of infected proportions in the cluster, and finer cluster values aggregate into coarser ones by a weighted sum. On the privacy side, the paper claims that a bounded Gaussian randomizer applied to each entity's local aggregated ERN vector, followed by a shuffler at each cluster, yields $(\\epsilon,\\delta)$-differential privacy for the cluster distributed ERN matrix, with explicit formulas for the mean and variance of the private values.","pith_inferences":["The threshold identity suggests a decentralized control rule: each region could push its own $\\bar R^t_i$ below 1 by reducing contact rates or shortening infection windows, without needing a central optimizer; the paper lists control as future work, so this is an extrapolation.","The privacy guarantee would be auditable if a calibration study computed the actual maximum $\\ell^2$ change in $\\zeta_i$ caused by adding or removing one person's mobility record through the paper's transmission-rate estimator, rather than asserting the $k=10^{-5}$ bound.","A streaming deployment that publishes private cluster ERNs on every day would need a composition analysis, because the paper analyzes a single release; repeated releases would consume the privacy budget over time.","The cluster threshold is defined with respect to a chosen partition and a weighted sum of infected proportions, so CERN values from different partitions should not be interpreted as the same epidemic quantity."],"forward_implications":["A local authority can tell whether its own region is expanding, contracting, or flat by comparing its local effective reproduction number to 1, using only its own transmission row and public infection proportions.","Cluster effective reproduction numbers can be aggregated hierarchically from local numbers, so county-, state-, and national-level monitors can follow the same outbreak at different resolutions without pooling raw mobility data.","If every local ERN is below 1, the network-level effective reproduction number is below 1, so local thresholds alone certify that the overall epidemic is declining.","The private cluster distributed ERNs preserve threshold-scale information: in the paper's COVID-19 mobility experiment, estimates at $\\epsilon=1$ track the nonprivate values with errors of roughly 5 to 9 percent.","The bounded Gaussian mechanism leaves zero entries at zero, so privacy noise does not fabricate transmission channels that do not exist in the network."],"supporting_citations":[{"why":"Supplies the network SIS/SIR model and the spectral-radius threshold conditions against which the local numbers are compared.","marker":"[7]"},{"why":"Supplies the network-level reproduction-number theory and equilibrium analysis that Theorem 2 connects to local ERNs.","marker":"[8]"},{"why":"Provides the flux-matrix method that turns mobility records into transmission rates in the experiments.","marker":"[16]"},{"why":"Defines differential privacy and the central-model guarantee used for the randomized outputs.","marker":"[23]"},{"why":"Defines next-generation matrices, which the paper identifies with the local distributed basic reproduction number matrix.","marker":"[31]"},{"why":"Provides the bounded Gaussian mechanism and its privacy-calibration optimization used as the local randomizer.","marker":"[42]"},{"why":"Provides the shuffle-model amplification bound used in Lemma 2 for cluster-level differential privacy.","marker":"[43]"},{"why":"Supplies the mobility dataset used to compute transmission rates for the real-world demonstration.","marker":"[44]"},{"why":"Supplies the daily COVID-19 case counts used to compute infected and susceptible proportions in the experiments.","marker":"[45]"}],"fun_headline_variants":["Private epidemic spread: local R0 still works","Differential privacy doesn't break local thresholds","Per-node R0 survives privacy noise","Epidemic thresholds from private data","Local spread rates, privacy preserved"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The privacy guarantee rests on the asserted bound that a single person's mobility record changes any local aggregated reproduction-number vector by at most $k=10^{-5}$ in $\\ell^2$ distance, a value the paper states but does not derive from the transmission-rate estimation procedure.","fun_headline_variants_meta":{"raw":{"variants":["Private epidemic spread: local R0 still works","Differential privacy doesn't break local thresholds","Per-node R0 survives privacy noise","Epidemic thresholds from private data","Local spread rates, privacy preserved"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1468,"prompt_tokens":1022,"completion_tokens":446,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":383}},"tokens_in":638,"tokens_out":446,"duration_ms":5203,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T22:12:04.527485+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the local aggregated ERN vector $\\zeta_i$ after adding and after removing one individual's daily visits in the Section V mobility data, using the same transmission-rate estimator; if the largest $\\ell^2$ change exceeds $k=10^{-5}$, the differential-privacy guarantee claimed for interaction frequencies does not hold on that data.","supporting_citations":[{"cited_title":"On the dynamics of deterministic epidemic propagation over networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the network SIS/SIR model and the spectral-radius threshold conditions against which the local numbers are compared."},{"cited_title":"Modeling, estimation, and analysis of epidemics over networks: An overview,","cited_arxiv_id":null,"evidence_quote":"Supplies the network-level reproduction-number theory and equilibrium analysis that Theorem 2 connects to local ERNs."},{"cited_title":"A high-resolution flux-matrix model describes the spread of diseases in a spatial network and the effect of mitigation strategies,","cited_arxiv_id":null,"evidence_quote":"Provides the flux-matrix method that turns mobility records into transmission rates in the experiments."},{"cited_title":"The algorithmic foundations of differential privacy,","cited_arxiv_id":null,"evidence_quote":"Defines differential privacy and the central-model guarantee used for the randomized outputs."},{"cited_title":"The construction of next-generation matrices for compartmental epidemic models,","cited_arxiv_id":null,"evidence_quote":"Defines next-generation matrices, which the paper identifies with the local distributed basic reproduction number matrix."},{"cited_title":"The bounded Gaussian mechanism for differen- tial privacy,","cited_arxiv_id":null,"evidence_quote":"Provides the bounded Gaussian mechanism and its privacy-calibration optimization used as the local randomizer."},{"cited_title":"Distributed differential privacy via shuffling,","cited_arxiv_id":null,"evidence_quote":"Provides the shuffle-model amplification bound used in Lemma 2 for cluster-level differential privacy."},{"cited_title":"SafeGraph Social Distancing Metrics,","cited_arxiv_id":null,"evidence_quote":"Supplies the mobility dataset used to compute transmission rates for the real-world demonstration."},{"cited_title":"An interactive web-based dashboard to track COVID-19 in real time,","cited_arxiv_id":null,"evidence_quote":"Supplies the daily COVID-19 case counts used to compute infected and susceptible proportions in the experiments."}],"review_version":1}